Block-based compressed sensing of MR images using multi-rate deep learning approach

Abstract

Deep learning (DL) models are highly research-oriented field in image compressive sensing in the recent studies. In compressive sensing theory, a signal is efficiently reconstructed from very small and limited number of measurements. Block-based compressive sensing is most promising and lenient compressive sensing (CS) approach mostly used to process large-sized videos and images: exploit low computational complexity and requires less memory. In block-based compressive sensing, a number of deep models are needed to train with each corresponding to different sampling rate. Compressive sensing performance is highly degraded through allocating low sampling rates to various blocks within same image or video frames. In this work, we proposed multi-rate method using deep neural networks for block-based compressive sensing of magnetic resonance images with performance that greatly outperforms existing state-of-the-art methods. The proposed approach is capable in smart allocation of exclusive sampling rate for each block within image, based on the image information and removing blocking artifacts in reconstructed MRI images. Each image block is separately sampled and reconstructed with different sampling rate and reassembled into a single image based on inter-correlation between blocks, to remove blocking artifacts. The proposed method surpasses the current state-of-the-arts in terms of reconstruction speed, reconstruction error, low computational complexity, and certain evaluation metrics such as peak signal-to-noise ratio (PSNR), structural similarity (SSIM), feature similarity (FSIM), and relative l2-norm error (RLNE).